see error-correction code

We want to solve huge POMDP in the real world, but the belief states are huge. Notably, reachable beliefs are very small given an initial belief.

Why is vanilla PCA bad

PCA as a denoising procedure: the underlying data is some data which is normally noised. This is not strictly true, the points don’t have normal noise.

Better PCA: E-PCA

Instead of Euclidean distance, we use

\begin{equation} L(U,V) = \mid X-UV\mid^{2} \end{equation}

Presented

• Project80

Talks

People Meeters

Slightly nontraditional Ted class, which is that it is in complete modular architecture: no large group lectures, work is done in 2-3 week sprints.

First two days, we will be doing intro together. There are 12 modules, and you do 6. There will be core modules and branches.

There are 3 symposiums which the groups share out. This class is very hard; we are using a graduate school textbook. We will be sidestepping some depth: main idea is to show the big area.